Built-in judges

Built-in judges are read-only. To change a prompt, choose Duplicate and edit the copy. Score judges pass when the score reaches the threshold shown. Scores run from 0 to 1, and higher is always better.

Judge	What it checks	Level	Result
Correctness	Whether the answer is factually and logically right. Uses a reference answer when you have one.	Span, trace, session	Pass / fail
Toxicity & Safety	Hateful, harassing, violent, sexual, or unsafe content in the output.	Span, trace, session	Pass / fail
PII Leakage	Personal data that should not have been shared.	Span, trace, session	Pass / fail
Bias	Different treatment of people based on protected traits.	Span, trace, session	Pass / fail
Prompt Injection	Whether the input tries to override the agent’s instructions.	Span, trace, session	Pass / fail
Conversation Quality	Whether the user probably left satisfied.	Session	Pass / fail
User Intent Change	Whether the user’s goal stayed the same. A change is information, not a failure.	Session	Descriptive
Completeness	Whether every part of the question was answered.	Span, trace, session	Score, passes at 0.7
Answer Relevance	Whether the answer engages with what was asked, right or wrong.	Span, trace, session	Score, passes at 0.7
Coherence	Whether the answer is consistent and well organised.	Span, trace, session	Score, passes at 0.7
Conciseness	Whether the answer is brief without dropping what was needed.	Span, trace, session	Score, passes at 0.6
Agent Flow	Whether the agent’s steps were logical and well ordered.	Session	Score, passes at 0.6
Reasoning Coherence	Whether the stated reasoning matches the action taken.	Tool and handoff spans	Score, passes at 0.7
Faithfulness (RAG)	Whether each claim is supported by the retrieved documents.	Trace, session	Score, passes at 0.8
Context Relevance (RAG)	Whether the retrieved documents were useful.	Trace, session	Score, passes at 0.6
Chunk Attribution (RAG)	How much of the retrieved content shaped the answer.	Trace, session	Score, passes at 0.5
Refusal Correctness	Whether the agent refused when it should and answered when it should.	Turn	Category
Tone	Labels the tone: neutral, friendly, professional, empathetic, and others. Never fails.	Span	Category

Which spans a text judge reads

Judges that grade text (Correctness, Completeness, Toxicity, PII Leakage, and similar) apply to agent, chain, LLM, and model spans. A tool call or retrieval has no answer to grade, so it reads Not eligible. Reasoning Coherence applies to tool and handoff spans. Trace and session judges read the whole trace or session.

RAG judges need retrieved documents

If a trace has no retrieved documents, a RAG judge’s cell reads Missing input. This is expected, not a failure.

Result types

Type	What you see
Pass / fail	A boolean verdict.
Score	A percentage from 0 to 100% that passes at its threshold. Green means it passed.
Category	A label, such as a tone. Each label can carry a score and pass or fail.
Descriptive	Information only, never a failure.

Pick a built-in judge

Visit the Judges page in BotGauge to enable any of the 18 judges and start scoring your traces.

Open the Judges page →

Next steps

To change what a judge does, duplicate it and edit the copy. To change how often or when it runs, see Rules and settings.

Get started

Judges

Results

Reference

Which spans a text judge reads

RAG judges need retrieved documents

Result types

Next steps

​Which spans a text judge reads

​RAG judges need retrieved documents

​Result types

​Next steps

Which spans a text judge reads

RAG judges need retrieved documents

Result types

Next steps